Pixel-level detection and measurement of concrete crack using faster region-based convolutional neural network and morphological feature extraction

Crack characteristics are important indicator reflecting the safety status of concrete structures. Current pixel-level crack detection methods generally used several semantic segmentation networks. However, those semantic segmentation network-based methods need expensive pixel-level annotation of training and test images. To overcome these problems, this paper proposed a pixel-level detection and measurement of concrete crack using a faster region-based convolutional neural network (faster R-CNN) and morphological feature extraction techniques. The faster R-CNN is trained on a database including 4861 crack images, and, consequently, records with 90.91% average precision (AP). The trained faster R-CNN is used to detect cracks from backgrounds of images, and then the morphological feature extraction techniques are used to segment pixel-level cracks and measure crack maximum widths and lengths. Comparative study is conducted to examine the performance of the proposed approach using a fully convolutional network (FCN)-based method. The results show that the proposed method substantiates quite performances and can indeed detect and measure concrete crack in realistic situations.